Where does AI go wrong in a law firm?

Usually in one of two places. A tool produces a confident answer with a citation nobody checked, and it ends up in a draft or a filing. Or someone pastes client information into a consumer chatbot because the firm never gave them an approved alternative.

In the Thomson Reuters Institute’s 2026 research on law firm AI strategy, clear information on accuracy and how to rigorously verify outputs was the number one barrier for lawyers slow to adopt. A firm can prevent both with better process, and that is where my work starts.

What do you build for law firms?

Research you can verify

I’m a historian by training, with a PhD from Northwestern, and my work has always turned on one question: what does the evidence actually show? My research workflows ask that question of AI. Each claim is separated out and labelled as confirmed in a source, inferred from one, or unsupported. Every citation is checked against the document it points to, and a source’s author and publisher are checked before anyone relies on it. Anything unconfirmed goes to a person, with the gap stated plainly.

Search across your firm’s own documents

A knowledge base built from your precedents, policies and past work product. Attorneys ask a question in plain language and get an answer with a link to the passage it came from. When your documents don’t answer the question, it says so instead of guessing. Before anyone relies on it, I test it against questions with known answers and measure how often it gets them right.

Training for attorneys and staff

Hands-on sessions built around each role’s real tasks, so people leave using the tools on their own matters instead of remembering a demo.

When a firm needs them, I also build intake agents, automations and internal dashboards as part of a larger engagement.

How do you work with our team?

I start by sitting with the people who do the work. My team and I meet attorneys and staff one on one, on their own screens and their own matters, and learn how each task actually gets done, including the steps nobody ever wrote down.

Those sessions do two jobs. People get help using Claude on work they have that week. And the notes and recordings become the material I build from: custom Claude tools, which are packaged instructions and checks that teach Claude how your team does a specific task, much like training a new hire with a procedure manual.

Then we go back to the same people, see what worked, and adjust. The tools fit how your firm works because they were built from watching it.

How does an engagement work?

1. Readiness and risk assessment. A fixed-price review of how your people use AI today, where the risks sit, and which workflows would pay off first. This first step is a fixed fee, from $5,500, and takes two to three weeks.
2. Policy and setup. A governance framework, a written AI use policy, and properly configured tools.
3. Working sessions with your team. One-on-one time with the people doing the work, on their own matters. Staff time: about an hour per person per week during the engagement, scheduled around their matters.
4. Build and refine. Custom Claude tools and research workflows built from those sessions, tested before rollout and adjusted after.
5. Training and handover. Role-based training, plus documentation your team can maintain without me.

Who have you done this for?

I’ve worked with an established personal injury firm and a boutique law firm in Houston on AI governance policy and firm-wide Claude training. Client work stays confidential, so I describe the type of work rather than the details.

Related: Claude training for teams, how I test Claude before anyone relies on it and my background.

He delivered what he promised. He was able to cleanly and clearly articulate what we were going to be doing. If there was ever an issue, he raised it in a way that was helpful and positive and timely. We didn’t get any of the sort of 11th-hour disasters that you might expect from an engineering project.

Rob Howard, CEO, Innovating with AI

Before consulting, I spent fifteen years teaching, including at Northwestern, Utah State and Manhattan College. That is where the training side of this work comes from.

Questions firms ask

Will our client data be used to train AI models?

Not on the plans I set up. I put firms on business or enterprise plans whose contracts say customer data isn’t used for training, and I confirm those terms are in place before any client information goes in.

Does this satisfy ABA Formal Opinion 512?

The setup and policy are designed around what the opinion covers: competence, confidentiality, communication with clients, supervision and fees. Whether your firm’s use meets your obligations is a judgment your lawyers make. My job is to give them a setup they can evaluate and a record that shows how it works.

Do we have to tell clients we use AI?

Opinion 512 calls for informed client consent before client information goes into AI tools that learn from what you enter. State bar guidance can go further: California’s, for example, asks firms to consider disclosure more broadly. Your AI use policy should settle this for your jurisdiction, and I’ll help you write that section.

Which AI tools do you work with?

Mainly Claude, from Anthropic. For firms already committed to Microsoft, I also work with Microsoft’s AI tools.

Are you a lawyer?

No. I’m a consultant and teacher. I don’t give legal advice, and I work alongside your attorneys.